It can depend on the manufacturer how well they actually measure their limits for guarantees. Usually the actual methodology is not documented in great detail, though they may have some type of general documentation. I've done the measurements for a few of these lines on datasheets. Basically, it is a guarantee by the manufacturer that the value will fall within the limits. Usually what's at stake is mainly the manufacturer's reputation -- there is often a clause somewhere that they will not be liable for other damages, though major customers do sometimes ask for big payments if they need to upgrade equipment in the field because a batch of parts was out of spec.
The simplest case is when the parameter is tested at production. In that case, we normally calculate the 6-sigma reproducibility and the 6-sigma repeatability errors and use those margins, often padding more to a round number. Reproducibility errors are based on testing on different testers and using different test boards/sites. Repeatability errors are based on repeating the test on the same part multiple times (often re-inserting into the socket).
The somewhat trickier parameters are ones that are not possible to measure at production test. Often this is because it takes too long to measure (e.g., sweeping temperature after packaging) or we don't have the right instruments on the tester. We have to measure these in the lab using slower methods (e.g., sweeping temperature 1 part at a time). Again, we often use 6-sigma margins on measurements, but this can sometimes be much more than you would think. For example, if the part has a type of internal comparison that runs once every clock period, we calculate how many times it would run at max speed over 10 years, and then extrapolate/calculate so that less than 1 part sold per million would fail once in that 10 years (1 ppm is actually a bit better than 6-sigma). That extreme example is really only if this type of failure would cause the part to become useless (e.g., requiring a power cycle).
The above is what we do for a "competitive" parameter -- one that distinguishes our part from the competition and is the reason somewhat would choose our part. If it's a parameter that is not very important, we sometimes use much larger margins just in case we did something wrong (e.g., missing some cause of variation). That way we don't get returns for parameters that we don't really care much about.
The other main strategy for specifying competitive parameters (like if the competitor is not using such a rigorous method) is to put a footnote that says exactly what is meant, e.g., "3-sigma in condition X with N=100." That's more of a marketing decision than an engineering one though.
